pytorch_cluster

• Fix backward compatibility issue with PyTorch < 1.12
• Remove torch.jit.script conversion of functions on module import

What's Changed

• Add batch_size argument for fps, knn, radius functions (https://github.com/rusty1s/pytorch_cluster/pull/175)
• Extend FPS with an extra ptr argument https://github.com/rusty1s/pytorch_cluster/pull/180)
• Add PyTorch 2.1.0 support (https://github.com/rusty1s/pytorch_cluster/pull/191)

New Contributors

• @piotrchmiel made their first contribution in https://github.com/rusty1s/pytorch_cluster/pull/175
• @cimadure made their first contribution in https://github.com/rusty1s/pytorch_cluster/pull/179
• @dsciebu made their first contribution in https://github.com/rusty1s/pytorch_cluster/pull/180
• @AgarwalSaurav made their first contribution in https://github.com/rusty1s/pytorch_cluster/pull/188

Full Changelog: https://github.com/rusty1s/pytorch_cluster/compare/1.6.1...1.6.2

• Added support for PyTorch 2.0
• Added support for return the indices of sampled edges in random_walk (#139)
• Added bf16 support for knn, radius and graclus (#144)
• Add safety checks on batch layout in nearest (#168)

• Improve calculation of num_nodes in random_walk (in case num_nodes=None) (#112)
• Fix knn/radius calculation for batches with zero-point examples
• Heavily improved efficiency of knn/radius calculation
• Half-precision support (#119)

• Reduced the size of shared library files
• CUDA wheels can now also operate on CPU-only devices
• Added parallelization strategies for CPU functionalities
• fps can now take in different ratios across different batched point sets, i.e. ratio can now be a torch.Tensor
• Fixed a bug in which radius computed slightly different results across CPU and CUDA versions

• PyTorch 1.7 wheels
• random_walk now supports q != 1 and p != 1 via rejection sampling

• PyTorch 1.6.0 wheels
• Fixed a bug in radius where the max_num_neighbors argument has been ignored

This release fixes some issues in the new knn and radius CPU implementations that have led to memory leaks. It is strongly recommended to update the package in case you are currently using torch-cluster==1.5.5.

• torch-cluster is now fully-jittable thanks to new implementations for knn and radius based on nanoflann rather than scipy.

Fixed a bug in the CUDA version of fps.

Fixes regarding proper usage of fps and radius.

This release introduces traceable torch-cluster operators using torch.jit.script. As a result, PyTorch 1.4 is now required to install this package. We also provide Python wheels for common OS/CUDA combinations (see README).

This release brings PyTorch 1.3 support to torch-cluster.

• PyTorch 1.2 update

• Random Walk Sampling API

Added flow argument to the radius_graph method.

• Added flow argument to the knn_graph method.

• Added NeighborSampler

• PyTorch 1.1.0

• multi-GPU support